Determinism Tier 1 follow-up: markdown audit reports embed a wall-clock timestamp

[mlieberman85]

Follow-up to #418 (Determinism Tier 1), which this gap escaped.

Summary

format_results_markdown in
packages/darnit/src/darnit/tools/audit.py embeds a wall-clock stamp
in every report:

**Generated At:** 2026-09-08T15:47:47.180106Z

Consequence: a markdown audit report can never be byte-for-byte
reproducible, even for an identical commit with an identical verdict
set. Two runs one second apart produce different bytes.

A second, smaller instance: pass_history[].duration_ms is
load-dependent. It reads 0 for fast file-existence checks in practice,
but nothing guarantees that.

How it surfaced

Feature 036's SC-002 required capturing a pre-feature output baseline
and diffing post-feature output against it byte-for-byte. Two
consecutive captures of the same commit did not match. Both fields had
to be scrubbed for the baseline to be stable at all -- see the scrub()
helper in tests/darnit/fixtures/error_class_baseline/capture_baseline.py
and its docstring, which documents both as out of scope for 036 and
points here.

#418's survey covered filesystem iteration order, wall-clock injection
into remediation templates, list ordering, and non-atomic writes. It
did not look at the report formatter, which is why this was missed.

Why it matters

Same argument #418 makes: a compliance tool whose report output drifts
run-to-run undermines the "same commit -> same result" premise that
makes an audit trustworthy. Concretely it also:

• Blocks any golden-file regression test of markdown output. Any future
  feature wanting to prove "my change is output-neutral" has to
  reinvent the same scrubbing.
• Makes report diffs noisy in review when an operator commits reports
  to a repo (some fleet operators do).

Scope

Decide between:

1. Drop the field. Simplest. The information is recoverable from
   the attestation timestamp or the file mtime for anyone who needs it.
2. Make it opt-in via a formatter argument, default off.
3. Make it injectable -- feature 036 already established this
   pattern for exactly this problem: RemediationExecutor gained a
   now_provider kwarg (defaulting to datetime.now) so tests can
   pin a fixed clock. Applying the same shape to
   format_results_markdown keeps production behavior while making
   the output testable.

Recommend (3) for consistency with the precedent, or (1) if nobody is
actually consuming the field.

Also decide whether duration_ms should be omitted from serialized
pass_history (it is diagnostic, not evidence) or kept and documented
as non-reproducible.

Non-goals

• Re-auditing the other formatters. JSON and SARIF do not embed
  timestamps at the per-result level; the attestation predicate has its
  own timestamp, which is legitimate and should stay.
• Any change to what the audit concludes. This is purely about output
  byte-stability.

Constitution alignment

• II. Conservative-by-Default: reproducible reports are a
  precondition for the report being usable as evidence.

Related

• Determinism Tier 1: normally deterministic operations must be deterministic across runs #418 (Determinism Tier 1) -- parent; closed by fix(determinism): Tier 1 -- deterministic ops must be deterministic (#418) #424 before this gap
  was found.
• Determinism Tier 2: network / side-effect failures must be distinguishable from real verdicts #419 / Determinism Tier 2: remediation must be atomic and idempotent under partial failure #420 / Determinism Tier 3: capture LLM reasoning and evidence trail in attestations, logs, and reports #421 -- the rest of the determinism set.
• Feature 036 (feat: distinguishable side-effect failures via error_class (feature 036) #435) -- surfaced this, and established the
  now_provider injection pattern option (3) would reuse.